A Cluster-Based Outlier Detection Scheme for Multivariate Data

Detection power of the squared Mahalanobis distance statistic is significantly reduced when several outliers exist within a multivariate dataset of interest. To overcome this masking effect, we propose a computer-intensive cluster-based approach that incorporates a reweighted version of Rousseeuw’s minimum covariance determinant method with a multi-step cluster-based algorithm that initially filters out potential masking points. Compared to the most robust procedures, simulation studies show that our new method is better for outlier detection. Additional real data comparisons are given. Supplementary materials for this article are available online.

当目标多变量数据集中存在多个异常值时，平方马氏距离（squared Mahalanobis distance）统计量的检测效能会显著降低。为克服该掩盖效应，本文提出一种计算机密集型的基于聚类的方法：该方法将鲁塞乌（Rousseeuw）最小协方差行列式（minimum covariance determinant, MCD）方法的加权版本，与可初步过滤潜在掩盖点的多步聚类算法相结合。相较于当前主流的稳健性检测流程，仿真实验结果表明，本文所提新方法在异常值检测任务中表现更优。此外还给出了真实数据集的对比分析结果。本文的补充材料可在线获取。